Raising: privatize lane-local scratch

[wsmoses]

A thread-private array inside the lane-batched parallel (mfem's real_t grad[3] accumulators in the dynamic 2D grad kernels) was modeled as one shared buffer. A store whose index map does not involve the lane axes looks uniform to the raising, so every lane read back lane zero's value — QuadratureInterpolator's tensor PhysDerivatives froze the qy axis (caught via a PJRT harness oracle against a numpy reference, minimized to a 40-line kernel).

Fix: a new preprocessing normalization gives such buffers one leading dimension per lane axis and indexes every access with the lane IVs (bounded to the batched extents, small-buffer-gated). Lit test included; the minimized kernel raises and computes bit-exact results.

Part of #2968.

🤖 Generated with Claude Code

https://claude.ai/code/session_016zErYp7upmqr4NHfhod9UD

[wsmoses]

Pushed two fixes found running the MFEM suite end-to-end:

• The 65k-element size budget was a silent miscompile: mfem's 3D register-blocked kernels keep per-lane register tensors of 24×24 lanes × 126 registers = 72k elements; the budget skipped privatization and the scratch raised as a single shared tensor, so every lane read the wrong values (VecDivPA 3D wrong by ~3e-2 while the statically-specialized instantiations — which never hit the budget — were exact). Budget raised to 4M elements. A follow-up should make a skipped privatization fail the strict raise loudly instead of miscompiling.
• The lit test now uses affine loops only, so it runs on this branch without the scf.for raising machinery (CI was red because of that).